Skip to content

Latest commit

 

History

History
37 lines (22 loc) · 1.2 KB

README.md

File metadata and controls

37 lines (22 loc) · 1.2 KB

Code for the paper: Motion Perception in Reinforcement Learning with Dynamic Objects

Artemij Amiranashvili, Alexey Dosovitskiy, Vladlen Koltun and Thomas Brox, CoRL 2018 (paper link).

Dependencies:

Installation:

Run pip3 install -e . in the gym and baselines directories.

Tasks:

  • Chaser2d-v2
  • Catcher2d-v2
  • Catcher3d-v1
  • KeepUp3d-v1
  • ChaserWithRandom4Backgrounds2d-v2
  • KeepUpHighMotionPenalty3d-v1

Running Examples:

Training Catcher3d-v1 with additional TinyFlowNet flow prediction input (replace LOG_DIR with path for logging):

python3 baselines/baselines/ppo2/run_mujoco_imvec.py --main_path LOG_DIR --env_id Catcher3d-v1 --add_flownet True --flownet_path networks/Catcher3d/

Training Chaser2d-v2 with image stack input:

python3 baselines/baselines/ppo2/run_mujoco_imvec.py --main_path LOG_DIR --stack_frames True

Training Chaser2d-v2 with additional image difference input:

python3 baselines/baselines/ppo2/run_mujoco_imvec.py --main_path LOG_DIR --stack_frames True --diff_frames True